14 December 2023, 2pm
Software tools to learn shape manifolds
3D shapes are best understood as samples on a curved manifold. Can we design robust and interpretable methods that “unfold” high-dimensional spaces of anatomical shapes with a minimal amount of training and parameter tuning? Going beyond tangent approximations of standard metrics, I will present fast numerical tools that enable non-linear manifold learning with real data. Notably, I will showcase a beta version of the “scikit-shapes” Python library and discuss the interaction of standard dimensionality reduction algorithms with the Wasserstein metric that is induced by optimal transport.